Standardizing Model Training and Feature Processing in Recommendation Systems

The talk, presented by senior Tencent engineer Liang Chao, focuses on the standardization of model training and usage processes within recommendation systems, where click‑through‑rate (CTR) estimation models are critical and complex.

In a typical recommendation pipeline, millions of items are filtered to a few dozen candidates through three steps: candidate generation using user, item, and context features; ranking with CTR or preference models; and final selection applying business rules and diversity constraints.

Key challenges include rapid feature iteration, maintaining consistency between online and offline feature handling, and supporting multiple model types. To address these, the team introduced a CSV‑driven feature framework that defines each feature’s name, type, position, and processing logic, allowing feature configuration, collection, and processing to be managed without extensive hand‑written code.

The framework standardizes feature types (int, sparse int, string, sparse string) inheriting from a base Feature class, ensuring identical serialization/deserialization for online logs and offline training samples. Feature processing follows three steps: filling, value/weight transformation, and joint vector transformation, illustrated with a tag feature example.

Training samples are generated in libsvm or sparse tensor formats, with the CSV converted to C++ header files and compiled into executables or shared objects that transform raw logs into training data. Dynamic compilation integrates custom TensorFlow operators when the CSV changes.

Additional components include feature monitoring (e.g., tag interest distribution), sample filtering and weighting to mitigate bot traffic, and a full system flow that ties feature configuration, code generation, online prediction, offline training, and re‑ranking together.

Overall, the CSV‑based standardization dramatically reduces manual coding, improves efficiency, and minimizes bugs in the feature engineering and model iteration lifecycle of recommendation systems.